A Calcined-Clay Mix-Design Study That Did Not Eat the Semester
The lab and the question
A graduate research group at a midwestern university runs a small but well-equipped concrete laboratory: a vane rheometer, a parallel-plate rheometer, a calorimeter, a small printing rig that the lab built itself for layered-extrusion specimens, and a programme of hardened-property tests routine enough that the senior students can run them without the advisor present. The group has a steady pipeline of mix-design and rheology questions, most of them framed around supplementary cementitious materials and the embodied-carbon profile of the resulting blends.
The question for this semester is whether a Limestone Calcined Clay Cement (LC3) formulation built around regionally sourced calcined kaolinitic clay can match the printability window of a conventional Type IL plus fly ash blend at meaningful clinker substitution levels, and whether the surface chemistry of the regionally-sourced calcined clay drives a measurable difference in static-yield-stress evolution compared to published European calcined-clay datasets. The hypothesis is that it does. The publication target is one of the cement-and-concrete journals.
The advisor has had this question in the queue for two years. It has not moved because the lab keeps running out of semester before it gets to the question itself.
What ate the previous attempts
The structural problem with semester-scale 3DCP mix-design research is upstream of the research question. To test a hypothesis about how a specific calcined clay drives static yield stress evolution, the lab first has to land a credible printable mortar that uses that calcined clay at all. That is not a test of the hypothesis; it is the precondition for the test. Landing it on locally-sourced inputs typically takes a multi-week gradation enumeration: trial dry blends against a Modified Andreasen and Andersen target curve, paste-volume sweeps to find the workable range, a superplasticizer compatibility screen because high-range water reducer interactions with metakaolin-rich systems differ from interactions with conventional supplementary cementitious materials, a viscosity-modifying-admixture screen, and a small set of bench prints to confirm the mortar even leaves the nozzle as a continuous filament before any of the actual rheology measurements get taken.
In the lab’s previous semester-long attempts at this question, the gradation-and-compatibility enumeration consumed most of the available bench time. The publishable measurements got run on a smaller candidate set than the advisor had intended, and the variance band on the static-yield-stress evolution data turned out to be wide enough that the comparison against the European calcined-clay datasets was not statistically clean. The paper draft sat for a year while the next graduate student was rotated in to redo the compatibility screen. That is a real loss: not just the project’s calendar, but the research question itself, which is interesting precisely because the field’s published calcined-clay 3DCP datasets are still thin and a careful comparison would be a contribution to the literature.
This is the canonical academic 3DCP problem. Bench time at the lab scale is the binding constraint, and the bench time is being spent on enumeration rather than on the question.
What CEMFORGE compressed
The lab adopted the CEMFORGE platform for the gradation-and-compatibility stage. The platform’s particle-packing module validates candidate dry blends against the Modified Andreasen and Andersen target curve as a hard physical gate before any rheological prediction is made; if a candidate skeleton cannot close volumetrically against the target distribution at the lab’s specified bounds, the engine rejects it rather than pass it forward. That alone removed a meaningful fraction of the candidates the lab would otherwise have bench-trialed, and the rejected candidates carry the reason for rejection (insufficient fines fraction, too-coarse skeleton, predicted excess water demand) in a form the graduate student can read against their materials log.
The rheological prediction layer narrowed the surviving candidate space further. CEMFORGE’s stacked machine-learning ensemble takes a candidate gradation, a chemistry, and a process window (target nozzle diameter, target layer time, ambient conditions) and emits predicted dynamic and static yield stress trajectories, predicted printability-window open and close times, and a calibrated training-coverage indicator that flags whether the candidate sits inside the trained envelope or in a sparser region of the chemistry space. For the regionally-sourced calcined clay specifically, the coverage indicator was honest: it flagged the platy-particle morphology of the local kaolinitic source as under-represented in the training corpus relative to the more spherical-fly-ash-dominant published literature. That is exactly the right output for an academic user. It tells the student where the model is interpolating and where they should expect the bench data to teach the model rather than the model to predict the bench data.
The candidate set the lab actually bench-trialed, after CEMFORGE’s pre-screen, was a small fraction of what the previous semester’s enumeration had attempted, and every candidate on the bench was there because the screen had a defensible reason for it. The bench prints closed quickly. The compatibility surprises came down to a small number of admixture-chemistry questions that the model had explicitly flagged as under-confident, which is the failure mode an academic user can absorb productively because it is informative and the lab learns from it.
What the lab did with the recovered time
The lab spent most of the semester on the actual research question: the static-yield-stress evolution of the regionally-sourced LC3 formulation across a controlled range of metakaolin-fraction and superplasticizer-dosage variations, with rest-time intervals fine enough to fit a credible structural-buildup-rate curve to each variant. The rheometer data set is dense enough that the comparison against the published European calcined-clay datasets is statistically meaningful, and the surface-chemistry interpretation the advisor wanted to make is supportable on the data the lab actually collected rather than backfilled from a literature aggregate.
The publication is in revision at the time of writing. The variance band on the static-yield-stress evolution is the band the advisor had hoped for two years ago, narrow enough to draw the comparison clean. The advisor’s view, which the case-study author has paraphrased rather than quoted, is that what changed was not the lab’s measurement quality but the fraction of the lab’s semester that was available for measurement at all.
The dataset, and why it leaves the lab in a useful shape
The lab elected to capture the project’s records in the Open3DCP schema, the Apache 2.0-licensed open data format for 3DCP mix-design records. CEMFORGE’s export produces an Open3DCP-conformant record per specimen, with the chemistry, gradation, process parameters, fresh-state rheology, and hardened-state properties as columns the schema recognizes. This is not academic housekeeping. It is the difference between a dataset that supports the lab’s paper and is then archived in a format only the next student in the lab can read, and a dataset that supports the lab’s paper and is also FAIR-compliant: findable, accessible, interoperable, reusable across the field. A subsequent group studying the same calcined-clay class on different equipment can pick up the lab’s Open3DCP file and either cross-validate against it or extend it without bilateral integration.
For a research group whose downstream funding cycle increasingly asks for a data-management plan, this is a practical advantage rather than a philosophical one. The data-management plan is, effectively, written by adopting the schema. The dataset is publishable to Zenodo or an institutional repository under an Apache 2.0 or CC-BY license without further restructuring. The paper’s supplementary information is the dataset itself, in a format the editor’s data-availability requirement recognizes.
The Open3DCP schema is documented at open3dcp.org and archived at Zenodo concept DOI 10.5281/zenodo.19647471. The schema is the data contract; the lab’s records are now a legal contributor to that contract.
What the lab did not get from CEMFORGE
CEMFORGE did not run the rheometer. It did not cast the specimens or interpret the surface-chemistry data or write the paper. None of that was the point. The point was that the lab’s bench instruments are productive only as far as the lab’s semester runs, and the previous semesters had been running out before the actual research question came up. Compressing the enumeration phase of mix design lets the academic loop close on the research question rather than on the precondition for the research question.
The pillar that grounds CEMFORGE’s engineering substrate is the 3D concrete printing mix design review on the CEMFORGE site. The methodological frame for the closed loop the lab effectively walked is described in the our-approach pillar on Sunnyday Technologies. The lab’s contribution to the literature, when the paper publishes, will sit upstream of both: it is the kind of careful, locally-grounded measurement work that the field’s published 3DCP datasets are short of, and it is the kind of work that benefits most directly from a tool that does not consume the lab’s bench time on the precondition.
What this case is not
This is not a claim that CEMFORGE replaces the lab’s expertise. The expertise is in the question, the experimental design, the instrument time, and the interpretation; CEMFORGE has none of those. This is also not a claim that physics-informed plus ML prediction is a substitute for measurement. The training-coverage flag on the regionally-sourced calcined clay is the explicit acknowledgement that the lab’s measurement is what teaches the model, not the other way around. What CEMFORGE replaces is the part of the lab’s semester that is enumeration: gradation-and-compatibility screening that has to happen before the research question is testable, and that has historically been done by hand because that is what bench labs have always done.
The change that lets a graduate student spend their semester on the research question is, in aggregate, the change that lets the field’s published 3DCP datasets become as dense as the field’s hardware capacity already implies they should be. Conventional bench labs with conventional bench discipline have been doing the right kind of work for fifteen years; what has been missing is a way to recover the share of their time that is enumeration overhead. That is what this case is about.
How to engage
The CEMFORGE platform is available as a self-service subscription at cemforge.ai, with academic licensing terms that include the Open3DCP-conformant export used in this case. Research engagements that need a deeper consulting touch (custom training-coverage analysis on a non-standard SCM, support for a specific journal’s data-availability requirement, or a structured workshop with the lab on integrating CEMFORGE into an existing experimental-design workflow) are scoped under CEMFORGE. For inquiries: info@sunn3d.com.
Representative case study for engagement-modeling. The research group, university, and clay source described are composites; the methodological detail reflects the workflow CEMFORGE supports today and the kind of mix-design study that is publishable in current cement-and-concrete journals. No specific lab, advisor, or institution is disclosed.